Untitled
Sri Seshadri
3/19/2020
Objective
- Forecast next 5 years of US Sales.
- Quantify the likelihood of sales surpassing 1.8 trillion.
- Assess if forecasting sales by region would help the overall sales forecast.
Data
| YearMonth |
1992 Jan to 2019 Dec |
| RecessionYears |
1, during the years 2001,2008 and 2009; 0 otherwise |
| RecessionCorrection |
1, during the years 2001,2008,2009 and 2015-16;0 otherwise |
| Sales |
Sales in millions (USD) |
| Sales_in_trillions |
Sales in trillions (USD) |
Total Business Sales in the US
Seasonality
Modeling issue 1: What should be the training and validation data?
Models trained with 2010-14 data would not learn the recession effect.
Demonstrate Issue 1
Demonstrate Issue 1 (contd)
What if the training data was to include 2015 data and the trained models be validated against 2016-19 data?
Modeling issue 2
Even if the models learnt the behavior at recession periods …
How do we know future recession years in the forecast horizon?
Solution options to modeling issue 1
- Leverage pre great recession data and its similarity to post great recession data.
- Asuume 2015 was a recession year and train models and test on 2016-2019.
Use Recession/Correction years as indicatior variables in each of the soluton strategy. Ignore great-recession data for model training.
Leverage pre-great recession data
Models trained with data from 1999-2003 when applied to 2015 data without retraining sufferes level/intercept mis-specification. The mis-specification is handled by adding the difference of sales in Dec 2003 and Dec 2014 to the forecast.
Leverage pre-great recession data
Models trained with data from 1999-2003 when applied to 2015 data without retraining sufferes level/intercept mis-specification. The mis-specification is handled by adding the difference of sales in Dec 2003 and Dec 2014 to the forecast.
Composite model
- Level adjusted ETS with ARIMA = LEARIMA ??
- An ETS model with dampening trend (trained on 1992-2001 data) to be used when a recession year is forecasted and ARIMA model on growth years.
- Here 2016 is forecasted to be recession year and 2015,2017-19 are deemed growth years.
- The composite model forecasts 2015 with ARIMA, 2016 with level adjusted ETS models and 2016 with ARIMA model.
Models trained on pre-great recession data
ˆ
Accuracy metrics
|
.model
|
.type
|
ME
|
RMSE
|
MAE
|
MPE
|
MAPE
|
MASE
|
ACF1
|
|
LMArima
|
Test
|
-0.034
|
0.071
|
0.060
|
-2.792
|
4.533
|
1.036
|
0.655
|
|
composite
|
Test
|
-0.105
|
0.112
|
0.105
|
-7.737
|
7.737
|
1.818
|
0.467
|
|
tslm
|
Test
|
-0.164
|
0.173
|
0.164
|
-12.332
|
12.332
|
2.846
|
0.724
|
Assuming 2015 as recession year…
Accuracy metrics
|
.model
|
.type
|
ME
|
RMSE
|
MAE
|
MPE
|
MAPE
|
MASE
|
ACF1
|
|
composite
|
Test
|
-0.029
|
0.054
|
0.042
|
-2.092
|
3.041
|
0.736
|
0.506
|
|
LMArima
|
Test
|
0.182
|
0.214
|
0.184
|
12.764
|
12.942
|
3.233
|
0.837
|
|
tslm
|
Test
|
-0.242
|
0.252
|
0.242
|
-17.384
|
17.384
|
4.244
|
0.788
|
Solution to issue 2: How do we forecast future recession
GDP output gap is the difference between GDP output and potential. Positive output gap indicates overheated market.
Educated guess that thre consecutive years of overheated market results in a recession the following year
Point forecasts for 2020-2024
Assumes 2020 is a recssion year
Likelihood for sales to surpass 1.8 trillion at least once
- Simulate 5 year forecasts based on bootstrapped validation errors.
- Simulate 5 year forecasts by modeling shocked time series data of sales.
Above methods simulate all possible recession scenarios (assuming utmost 1 year of recession is possible)
Simulation based on bootstrapped test errors
The chance of sales surpassing 1.8 trillion US dollars is 44%
Simulation based on shocked time series data of sales
The chance of sales surpassing 1.8 trillion US dollars is 40%